Artificial intelligence-based classification of Spitz tumors

Background and AimsAccurate differentiation of cardiac amyloidosis (CA) from phenotypic mimics remains challenging using current clinical and echocardiographic techniques. The accuracy of a novel artificial intelligence (AI) screening algorithm for echocardiography-based CA detection was assessed.MethodsUtilizing a multisite, multiethnic dataset (n = 2612, 52% CA), a convolutional neural network was trained to differentiate CA from phenotypic controls using transthoracic apical four-chamber video clips. External validation was conducted globally across 18 sites including 597 CA cases and 2122 controls. Classification accuracy was assessed on the entire external validation dataset, and subgroup analyses were performed both on technetium pyrophosphate scintigraphy referrals, and individuals matched for age, sex, and wall thickness. Model accuracy was also compared with the transthyretin CA score and the increased wall thickness score within a subset of older heart failure with preserved ejection fraction patients with increased wall thickness.ResultsCardiac amyloidosis patients and controls displayed similar age, sex, race, and comorbidities. After the removal of uncertain AI predictions (13%), model discrimination and classification were excellent for the entire external validation dataset [area under the receiver operating characteristic curve (AUROC) 0.93, sensitivity 85%, specificity 93%], irrespective of CA subtype (sensitivity: light-chain = 84%, wild-type transthyretin = 85%, and hereditary transthyretin = 86%). Performance was maintained in subgroup analysis in patients clinically referred for technetium pyrophosphate scintigraphy imaging (AUROC 0.86, sensitivity 77%, specificity 86%) and matched patients (AUROC 0.92, sensitivity 84%, specificity 91%). The AI model (AUROC 0.93) also outperformed transthyretin CA score (AUROC 0.73) and increased wall thickness (AUROC 0.80) scores.ConclusionsThis AI screening model—using only an apical four-chamber view—effectively differentiated CA from other causes of increased left ventricular wall thickness.

Delirium is an acute syndrome characterized by fluctuating attention, cognitive impairment, and severe disorganization of behavior, which has been shown to affect up to 31% of patients in the intensive care unit (ICU). Early detection can enable timely interventions and improved health outcomes. While artificial intelligence (AI) models have shown great potential for ICU delirium prediction using structured electronic health records (EHR), most studies have either not leveraged state-of-the-art AI models, been limited to single-center cohorts, or relied on small datasets for development and validation. In this study, we introduce DeLLiriuM, a novel LLM-based delirium prediction model that utilizes EHR data from the first 24 hours of ICU admission to estimate a patient’s risk of developing delirium for the remainder of their ICU stay. We developed and validated DeLLiriuM using ICU admissions from 104,303 patients across 195 hospitals in three large databases: the eICU Collaborative Research Database, the Medical Information Mart for Intensive Care (MIMIC)-IV, and the University of Florida’s Integrated Data Repository. Our DeLLiriuM model achieved superior performance compared to all baseline models on the external validation set, measured by the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) metric. DeLLiriuM attained AUROC 82.4 (95% confidence interval 81.8–83.0) and AUPRC 11.8 (95% confidence interval 11.3–12.4) across 77,543 patients spanning 194 hospitals. Our approach of transforming structured EHR data into an unstructured text format, the primary data modality for LLMs, enables our DeLLiriuM model to capture clinical contextual information, resulting in improved predictive performance. To the best of our knowledge, DeLLiriuM is the first LLM-based delirium prediction tool for the ICU that utilizes structured EHR data with LLMs rather than clinical notes with LLMs or traditional structured feature representations used in AI models.

Rapid myocardial revascularization in patients with acute myocardial infarction (AMI) is essential to improve clinical outcomes. There is still room for improvement in the timely diagnosis of AMI. This study aimed to develop an artificial intelligence (AI) model using electrocardiograms (ECGs) for detecting AMI needing revascularization. A total of 723 389 ECGs from 300 627 patients in the derivation cohort at a single centre between 2013 and 2020, including 5872 patients with AMI (1.95%) who underwent revascularization, were used for model training and internal testing. A transformer-based deep learning model, initially trained on about one million unlabelled ECGs through self-supervised learning, was fine-tuned for AMI detection. The model's final performance was evaluated in the internal test and the external validation set. The external validation was conducted at an independent centre between 2002 and 2020 using 261 429 ECGs from 259 454 patients, including 1095 patients with AMI (0.42%). By integrating self-supervised learning to train the AI model, we enhanced the AMI detection performance, as demonstrated by an increase in the area under the receiver operating characteristic curve (AUROC) from 0.910 (95% CI, 0.904-0.915) to 0.968 (95% CI, 0.965-0.971) in the external validation set. For ST-elevation myocardial infarction and non-ST-elevation myocardial infarction detection, the AUROCs were 0.991 (95% CI, 0.989-0.993) and 0.947 (95% CI, 0.942-0.952) in the external validation set, respectively. This novel ECG-based AI model may be beneficial for the timely identification of patients with AMI needing revascularization.

ObjectiveTo develop and evaluate a deep learning-based artificial intelligence (AI) model for detecting skull fractures on plain radiographs in children.Materials and MethodsThis retrospective multi-center study consisted of a development dataset acquired from two hospitals (n = 149 and 264) and an external test set (n = 95) from a third hospital. Datasets included children with head trauma who underwent both skull radiography and cranial computed tomography (CT). The development dataset was split into training, tuning, and internal test sets in a ratio of 7:1:2. The reference standard for skull fracture was cranial CT. Two radiology residents, a pediatric radiologist, and two emergency physicians participated in a two-session observer study on an external test set with and without AI assistance. We obtained the area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity along with their 95% confidence intervals (CIs).ResultsThe AI model showed an AUROC of 0.922 (95% CI, 0.842–0.969) in the internal test set and 0.870 (95% CI, 0.785–0.930) in the external test set. The model had a sensitivity of 81.1% (95% CI, 64.8%–92.0%) and specificity of 91.3% (95% CI, 79.2%–97.6%) for the internal test set and 78.9% (95% CI, 54.4%–93.9%) and 88.2% (95% CI, 78.7%–94.4%), respectively, for the external test set. With the model’s assistance, significant AUROC improvement was observed in radiology residents (pooled results) and emergency physicians (pooled results) with the difference from reading without AI assistance of 0.094 (95% CI, 0.020–0.168; p = 0.012) and 0.069 (95% CI, 0.002–0.136; p = 0.043), respectively, but not in the pediatric radiologist with the difference of 0.008 (95% CI, -0.074–0.090; p = 0.850).ConclusionA deep learning-based AI model improved the performance of inexperienced radiologists and emergency physicians in diagnosing pediatric skull fractures on plain radiographs.

Artificial intelligence (AI) models can sense subjective affective states from facial images. Although recent psychological studies have indicated that dimensional affective states of valence and arousal are systematically associated with facial expressions, no AI models have been developed to estimate these affective states from facial images based on empirical data. We developed a recurrent neural network-based AI model to estimate subjective valence and arousal states from facial images. We trained our model using a database containing participant valence/arousal states and facial images. Leave-one-out cross-validation supported the validity of the model for predicting subjective valence and arousal states. We further validated the effectiveness of the model by analyzing a dataset containing participant valence/arousal ratings and facial videos. The model predicted second-by-second valence and arousal states, with prediction performance comparable to that of FaceReader, a commercial AI model that estimates dimensional affective states based on a different approach. We constructed a graphical user interface to show real-time affective valence and arousal states by analyzing facial video data. Our model is the first distributable AI model for sensing affective valence and arousal from facial images/videos to be developed based on an empirical database; we anticipate that it will have many practical uses, such as in mental health monitoring and marketing research.

Prognosis prediction is crucial for non-small cell lung cancer (NSCLC) treatment planning. While tumor hypoxia significantly impacts patient outcomes, identifying hypoxic genomic markers remains challenging. This study sought to identify hypoxic computed tomography (CT) radiomic features and create an artificial intelligence (AI) model for NSCLC through the integration of multi-modal data. In total, 452 NSCLC patients were enrolled in this study, including patients from The Second Affiliated Hospital of Soochow University (SC, n=112), The Cancer Genome Atlas (TCGA)-NSCLC dataset (n=74), the radiogenomics dataset (n=130), and the Gene Expression Omnibus (GEO) datasets (GSE19188: n=82, and GSE87340: n=54). Hypoxia status was classified using optimized cut-off values of hypoxia enrichment scores, which were calculated through single-sample gene set enrichment analysis (ssGSEA) of hypoxic genes. Radiomic features were extracted using three-dimensional (3D)-Slicer software. The least absolute shrinkage and selection operator (LASSO) algorithm was used to identify hypoxic CT radiomic features. A model named ssuBERT (semantic structured unit embedded in Bidirectional Encoder Representations from Transformers) was developed to analyze electronic health records (EHRs). An AI model for overall survival prediction was constructed by integrating CT radiomic features, ssuBERT features, and clinical data, and evaluated using five-fold cross-validation. Higher hypoxia levels were correlated with worse survival outcomes. Twenty-eight radiomic features showed significant discriminatory power in detecting hypoxia status with an area under the curve (AUC) of 0.8295. The ssuBERT model achieved a weighted accuracy of 0.945 in recognizing semantic structured units in EHRs. The EHR model exhibited superior predictive performance among the single-modal models with an AUC of 0.7662. However, the multi-modal AI model had the highest average AUC of 0.8449 and an F1 score of 0.7557. The AI model demonstrated potential in predicting NSCLC patient prognosis through multi-modal data integration, warranting further validation.

Democratizing artificial intelligence (AI) enables model development by clinicians with a lack of coding expertise, powerful computing resources, and large, well-labeled data sets. To determine whether resource-constrained clinicians can use self-training via automated machine learning (ML) and public data sets to design high-performing diabetic retinopathy classification models. This diagnostic quality improvement study was conducted from January 1, 2021, to December 31, 2021. A self-training method without coding was used on 2 public data sets with retinal images from patients in France (Messidor-2 [n = 1748]) and the UK and US (EyePACS [n = 58 689]) and externally validated on 1 data set with retinal images from patients of a private Egyptian medical retina clinic (Egypt [n = 210]). An AI model was trained to classify referable diabetic retinopathy as an exemplar use case. Messidor-2 images were assigned adjudicated labels available on Kaggle; 4 images were deemed ungradable and excluded, leaving 1744 images. A total of 300 images randomly selected from the EyePACS data set were independently relabeled by 3 blinded retina specialists using the International Classification of Diabetic Retinopathy protocol for diabetic retinopathy grade and diabetic macular edema presence; 19 images were deemed ungradable, leaving 281 images. Data analysis was performed from February 1 to February 28, 2021. Using public data sets, a teacher model was trained with labeled images using supervised learning. Next, the resulting predictions, termed pseudolabels, were used on an unlabeled public data set. Finally, a student model was trained with the existing labeled images and the additional pseudolabeled images. The analyzed metrics for the models included the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, and F1 score. The Fisher exact test was performed, and 2-tailed P values were calculated for failure case analysis. For the internal validation data sets, AUROC values for performance ranged from 0.886 to 0.939 for the teacher model and from 0.916 to 0.951 for the student model. For external validation of automated ML model performance, AUROC values and accuracy were 0.964 and 93.3% for the teacher model, 0.950 and 96.7% for the student model, and 0.890 and 94.3% for the manually coded bespoke model, respectively. These findings suggest that self-training using automated ML is an effective method to increase both model performance and generalizability while decreasing the need for costly expert labeling. This approach advances the democratization of AI by enabling clinicians without coding expertise or access to large, well-labeled private data sets to develop their own AI models.

e15161 Background: Analyzing genetic mutations is essential for personalized therapy decisions in oncology and beyond. In colorectal cancer (CRC), mutations in BRAF ( BRAFmut) render patients eligible for targeted treatment with BRAF inhibitors and immunotherapy. BRAFmut is routinely analyzed via genomic sequencing, which, however, is costly and does not provide diagnostic results immediately. Reliable detection of BRAFmut status from histological images using artificial intelligence (AI) could provide a more accessible and faster diagnostic alternative. Previous approaches to image-based BRAFmut analysis, however, failed to reach a critical level of accuracy on independent test cohorts, with reported area under the receiver operating characteristic curve (AUROC) values typically below 0.8. Methods: We hypothesized that combining tissue image analysis with clinical patient data could improve diagnostic accuracy for BRAFmut. Following our hypothesis, we developed a multi-modal AI model on n = 426 CRC cases from the TCGA database using BRAFmut data, H&E-stained tissue images, and information on patient parameters that are readily available in clinical practice and that are known to be associated with BRAFmut frequency. The model was based on a vision transformer architecture incorporating pathology-specific information using a pretrained embedding model for encoding tissue image features. Model performance was validated on an independent hold-out TCGA cohort of n = 107 samples and an additional external cohort of n = 104 CRC samples from the CPTAC database. Results: Our multi-modal AI model reached an AUROC of 0.85 ± 0.02 on the TCGA hold-out test set. Notably, on the second, external CPTAC dataset the predictive accuracy level was similar (AUROC of 0.83 ± 0.02) indicating the robustness of our approach to generalize to different datasets. Additional analyses showed that the significant performance boost over previous approaches was gained by the multi-modal analysis setup, as predicting BRAFmut based on either tissue, or clinical data alone led to lower accuracy values. Explainability analysis of AI results confirmed that relevant biological features on the histological slides, mainly tumor areas, were considered by the model for decision making. Conclusions: The combination of H&E images with clinically readily available patient parameters enabled us to develop a model predicting BRAFmut in CRC that surpasses previous approaches in diagnostic accuracy on independent evaluation cohorts. This indicates that combining histological tissue information with clinical parameters can increase the ability of AI models to predict genetic variants from tissue and more generally highlights the potential of multi-modal deep learning approaches to enhance predictive AI models in digital pathology.

A systematic review and meta-analysis of medical image-based artificial intelligence models for predicting metastasis in ovarian cancer.

There is growing evidence that the Spitz group of melanocytic neoplasms should be restricted to those harboring kinase receptor fusions and HRAS mutations/11p15 amplification. The presence of genomic alterations characteristic of conventional melanomas (BRAF and NRAS mutations) precludes a diagnosis of a Spitz neoplasm. It is often challenging to distinguish Spitz neoplasms from conventional melanomas with spitzoid morphology on histopathological grounds alone. We report a series of 70 consecutive melanocytic tumors in which targeted sequencing was indicated to distinguish Spitz from spitzoid neoplasms and to classify Spitz neoplasms along the biological spectrum. Final diagnoses incorporating molecular results included 12 conventional melanomas (nine of which with NRAS mutations), five Spitz melanomas, 35 atypical Spitz tumors, eight Spitz nevi, three melanocytic tumors with a MAP2K1 mutation, and seven desmoplastic Spitz nevi/tumors. There were significant discrepancies between initial diagnoses and final diagnoses after incorporating molecular results in 24 (34%) cases, including nine conventional melanomas favored to be Spitz neoplasms and nine Spitz neoplasms favored to be conventional melanomas. It is often not possible to reliably distinguish Spitz neoplasms from spitzoid melanocytic tumors without identifying their driver genomic alterations. Applying next-generation sequencing in diagnostically problematic tumors improves diagnostic accuracy.

BackgroundDeep learning has been widely used for glaucoma diagnosis. However, there is no clinically validated algorithm for glaucoma incidence and progression prediction. This study aims to develop a clinically feasible deep-learning system for predicting and stratifying the risk of glaucoma onset and progression based on color fundus photographs (CFPs), with clinical validation of performance in external population cohorts.MethodsWe established data sets of CFPs and visual fields collected from longitudinal cohorts. The mean follow-up duration was 3 to 5 years across the data sets. Artificial intelligence (AI) models were developed to predict future glaucoma incidence and progression based on the CFPs of 17,497 eyes in 9346 patients. The area under the receiver operating characteristic (AUROC) curve, sensitivity, and specificity of the AI models were calculated with reference to the labels provided by experienced ophthalmologists. Incidence and progression of glaucoma were determined based on longitudinal CFP images or visual fields, respectively.ResultsThe AI model to predict glaucoma incidence achieved an AUROC of 0.90 (0.81–0.99) in the validation set and demonstrated good generalizability, with AUROCs of 0.89 (0.83–0.95) and 0.88 (0.79–0.97) in external test sets 1 and 2, respectively. The AI model to predict glaucoma progression achieved an AUROC of 0.91 (0.88–0.94) in the validation set, and also demonstrated outstanding predictive performance with AUROCs of 0.87 (0.81–0.92) and 0.88 (0.83–0.94) in external test sets 1 and 2, respectively.ConclusionOur study demonstrates the feasibility of deep-learning algorithms in the early detection and prediction of glaucoma progression.FUNDINGNational Natural Science Foundation of China (NSFC); the High-level Hospital Construction Project, Zhongshan Ophthalmic Center, Sun Yat-sen University; the Science and Technology Program of Guangzhou, China (2021), the Science and Technology Development Fund (FDCT) of Macau, and FDCT-NSFC.

The revisit of the emergency department (ED) is a key indicator of emergency care quality. Various strategies have been proposed to reduce ED revisits, including the use of artificial intelligence (AI) models for prediction. However, AI model performance varies significantly, and its true predictive capability remains unclear. To address these gaps, the primary purpose of this study is to evaluate the performance of AI in predicting ED revisits through a meta-analysis. Specifically, this study aims to (1) Quantitatively assess the predictive performance of AI in ED revisit prediction and (2) Identify covariates contributing to between-study heterogeneity. A systematic search was conducted on December 31, 2024, across multiple electronic databases, including Scopus, SpringerLink, ScienceDirect, PubMed, Wiley, Sage, and Google Scholar, to identify relevant studies meeting the following criteria: (1) Utilized machine learning, deep learning, or artificial intelligence techniques to predict patient return visits to the ED, (2) Written in English, and (3) Peer-reviewed. Diagnostic accuracy was assessed using pooled sensitivity, specificity, and area under receiver operating characteristic curve (AUROC), while subgroup analysis explored factors contributing to heterogeneity. This meta-analysis included 20 articles, comprising 27 AI models. The summary estimates for ED revisit prediction were as follows: (1) Sensitivity: 0.56 (95% Confidence Interval [CI]: 0.44-0.67), (2) Specificity: 0.92 (95% CI: 0.86-0.96), and (3) AUROC: 0.81 (95% CI: 0.71-0.88). Subgroup analysis identified nationality, missing value-handling strategies, and specific disease samples as potential contributors to between-study heterogeneity. Future research should focus on improving missing value processing and using specific disease samples to enhance model reliability.

• Artificial intelligence (AI) models demonstrate consistently high diagnostic performance for intracranial aneurysm rupture status or rupture-related risk assessment across multiple metrics. • Diagnostic performance of human readers varies substantially according to expertise, with expert readers outperforming non-experts in sensitivity but showing similar overall accuracy. • Human–AI collaborative approaches show promising diagnostic performance in individual studies; however, current evidence is limited and highly dependent on integration strategy. • Considerable methodological heterogeneity, inconsistent outcome definitions, and limited head-to-head comparisons restrict definitive conclusions regarding comparative superiority. • Standardized definitions, transparent reporting, and multicenter prospective studies are required to clarify the optimal role of AI and Human–AI collaboration in aneurysm care. Accurate assessment of intracranial aneurysm rupture risk remains challenging, with conventional clinical scores showing limited predictive performance. Artificial intelligence (AI) has emerged as a potential adjunct to imaging-based risk stratification, yet the relative performance of AI, human experts, and Human–AI collaboration has not been systematically quantified. A systematic review and meta-analysis were conducted in accordance with PRISMA guidelines. PubMed, Scopus, Web of Science, and Embase were searched. Studies comparing AI-alone, human-alone, or Human–AI approaches for intracranial aneurysm rupture status or rupture-related risk assessment were included. The outcomes of interest were sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUC). Random-effects models were applied when appropriate, and methodological quality was assessed using the ROBINS-I tool. Seventeen studies were included. AI-alone models demonstrated consistently high diagnostic performance across sensitivity, accuracy, and AUC, although substantial heterogeneity was observed across methodologies and imaging modalities. Human reader sensitivity varied according to expertise, with expert readers achieving higher sensitivity than non-experts but similar overall accuracy. Human–AI collaborative approaches showed high diagnostic metrics in individual studies; however, the number of eligible studies was limited, and integration strategies were heterogeneous, precluding definitive pooled comparisons. Current evidence indicates that AI models can achieve high diagnostic performance in intracranial aneurysm rupture risk assessment and may serve as a valuable adjunct to clinical expertise. While Human–AI collaboration shows promise, available data remain exploratory and strategy-dependent. Standardized outcome definitions and multicenter prospective studies are required before firm conclusions regarding comparative effectiveness or clinical superiority can be established.

Anthracycline-induced cardiotoxicity is a major concern in childhood cancer survivors. Detecting subclinical cardiac dysfunction early is critical, but the accuracy of current diagnostic tools is uncertain. We performed a systematic review and meta-analysis of studies evaluating the prognostic accuracy of echocardiography, serum biomarkers, microRNAs (miRNAs), and artificial intelligence (AI) models in predicting chemotherapy-induced cardiotoxicity in pediatric patients. Seventy-three studies were included in the qualitative synthesis, and ten in the meta-analysis. Quality was assessed using the QUAPAS tool (CRD42023485629). AI-based models showed the highest pooled predictive performance area under the curve (AUC)~0.80, despite significant heterogeneity (I² = 93%). Global longitudinal strain (GLS) had moderate accuracy (AUC~0.72). Cardiac biomarkers like troponin and N-terminal pro-B-type natriuretic peptide (NT-proBNP) showed variable sensitivity and specificity, largely influenced by timing and thresholds. Preliminary evidence on miRNAs was promising but lacked standardization. Additional methods (e.g., cardiac MRI, cardiopulmonary exercise testing (CPET)) were excluded from meta-analysis due to methodological variability. No single modality reliably detects early cardiotoxicity. Promising tools like AI models and miRNAs need further validation. A multimodal diagnostic strategy combining imaging, biomarkers, and clinical data may be the most effective approach. Standardized definitions and protocols are urgently needed. Early detection of chemotherapy-induced cardiotoxicity in children remains unresolved. Multimodal assessment is most effective because no single test is sufficiently reliable. There are critical methodological gaps and heterogeneity that impede standardization in pediatric cardio-oncology. AI and microRNAs are promising but still unvalidated tools. The findings guide future clinical monitoring strategies and support standardized multimodal diagnostic algorithms.

Classical scoring tools such as the Bronchiectasis Severity Index (BSI) and FACED have shown good performance for predicting adverse outcomes, yet they were developed using European cohorts and may not fully align with Asian clinical characteristics. This study aimed to develop an Asian population–specific artificial intelligence (AI) model for predicting severe acute exacerbations (AEs) in bronchiectasis. A total of 492 patients with 1-year follow-up data from the Korean Multicenter Bronchiectasis Audit and Research Collaboration registry were analyzed. Severe AE was defined as an event requiring an emergency department visit or hospitalization due to worsening respiratory symptoms. Three AI models—extreme gradient boosting, logistic regression, and multilayer perceptron (MLP)—were trained and compared with classical scoring models. Severe AE occurred in 56 patients (11.4%). Among the models, the MLP demonstrated the best performance, with higher sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve compared with both classical scores. Shapley additive explanation (SHAP) analysis identified BSI, sputum characteristics, and histories of tuberculosis and pneumonia as key predictors. Although classical scores performed well, a population-specific AI model using local clinical data showed improved predictive performance and may support individualized risk assessment for Asian patients with bronchiectasis.